More than one-third of English-language webpages published since the launch of ChatGPT show measurable evidence of being generated or heavily edited by large language models, according to a study released by the Pew Research Center.
The analysis examined nearly 500,000 English-language web documents indexed across a five-year window in the Common Crawl archive, spanning periods before and after OpenAI introduced ChatGPT in November 2022. Using detection technology developed by Open Pangram, researchers evaluated the prevalence of synthetic text across the contemporary open web.
Domain Disparities and Crawl Filtering
In an unfiltered random sample of 10,000 webpages retrieved from a July 2026 Common Crawl snapshot, approximately 10% of total documents triggered classification thresholds for significant AI authorship. Because broad web crawls capture archival content published years prior to modern generative models, researchers isolated pages published specifically after November 2022. Within that post-ChatGPT cohort, 35% of all examined webpages exhibited substantial machine generation or automated revision.
The distribution of machine-authored content varies sharply across top-level domain categories:
- Commercial Domains (.com): Exhibited the highest concentration of synthetic text, appearing at roughly ten times the rate of academic or governmental domains.
- Non-Profit Registries (.org): Showed an intermediate adoption rate, with 4.6% of analyzed pages flagged as machine-written.
- Educational and Government Portals (.edu / .gov): Remained largely resistant to automated publishing, registering AI authorship rates of approximately 1%.
Mode Collapse and Stylistic Fingerprints
The proliferation of automated text across the web mirrors recent data from internet infrastructure provider Cloudflare, which confirmed earlier this year that automated bot traffic has surpassed human traffic across global routing networks.
Pew researchers also noted a measurable uptick in specific rhetorical patterns and formatting markers associated with frontier model outputs. These include disproportionate increases in Oxford commas, antithetical constructions (such as "it is not X, it is Y"), and specific punctuation distributions.
AI text detection engines rely heavily on "mode collapse" introduced during model alignment. While pre-trained base models retain wide lexical and structural variance that mimics human diversity, post-training techniques such as Reinforcement Learning from Human Feedback (RLHF) and safety steering compress probability distributions toward standardized phrasing. This concentration around narrow modes creates statistical regularities that classifiers detect, even in the absence of explicit cryptographic or sampling watermarks.
